AI Transport v0.5.0 Enables Durable Agent Execution with Steps Framework
Key Takeaways
- ▸Steps provide idempotent output handling: retried work replaces previous failed attempts instead of duplicating messages to users
- ▸Agent execution lifecycle can now span multiple processes under a single stable runId, enabling resilient deployment architectures
- ▸Output delivery persists on Ably channels, enabling seamless reconnections and multi-device synchronization without manual offset tracking
Summary
AI Transport v0.5.0 introduces first-class support for running AI agent turns within durable execution frameworks like Temporal and Vercel's Workflow Development Kit (WDK), while maintaining a single clean, resumable stream for all watching clients. The release addresses the challenge of running agents reliably when processes may crash, restart, or be reclaimed mid-execution.
The core innovation is the introduction of Steps—atomic, re-attemptable units of output that prevent duplicate messages when a durable framework retries failed work. Each Step carries a stable stepId that ensures retried work replaces previous failed attempts rather than duplicating output to users. This allows developers to safely integrate agent execution into production infrastructure without manually handling idempotency keys or deduplication logic.
V0.5.0 extends the library's previous capability (v0.4.0's agent history hydration from custom databases) to solve the second half of reliable agent execution: persisting and delivering the agent's streamed output reliably across clients, reconnections, and multiple devices. The release maintains backward compatibility—existing single-process agents using run.pipe() continue to work without changes, with output automatically bracketed in implicit steps.
- Backward compatible—existing agents using run.pipe() work without changes; explicit stepId only needed when retries are possible
- Solves the gap between durable execution frameworks (Temporal, Vercel WDK) and reliable client delivery in production AI applications
